BH28 trades without me pressing a button. Every trading day the system fetches the data itself, checks it, applies the strategy, puts the order through the risk gate, executes it on a simulation account and then compares what actually happened. I follow along through alerts and reports. No real money is involved, and I do not mention returns here.
BH28 in brief
- Status: built and completed, running autonomously on a simulation account.
- 32 n8n workflows for scheduling, handover, monitoring and reporting around the core.
- Python core: strategy and risk rules are deterministic. The same input always gives the same outcome.
- PostgreSQL stores the state and an audit of every decision, including trades that did not go ahead.
- ChatGPT and Claude help with review, analysis and reporting, without authority over orders.
- Three separate routes: futures on a funded account first; futures on my own account and an options strategy as separate research routes, each with its own rules and risk budget.
A system that trades on its own must above all be good at saying no
A trading strategy on paper is easy to follow. In software it gets tricky: duplicate orders, delayed data, an order that is only partly filled, a connection that drops, or a system that assumes a position is already closed when it is not. That is why in BH28 every step can stop the rest.
If a fact is missing, or data contradict each other, the status becomes “unknown” and the chain stops. Unknown never quietly turns green. A trade that is skipped because the data do not add up is a good outcome.
A trading day, as the system reads it
An example of the kind of rule BH28 carries out. The range of the first half hour sets a buy stop above it and a sell stop below it. If one is triggered, the other is cancelled. Straight after the fill, the protection must demonstrably be in place: stop-loss and take-profit. If the system cannot confirm that, or the connection drops, it first reconciles with the broker. An order that may already have arrived is never blindly sent again.
The risk gate
Every order has to pass the same checks: is the market data current and complete, does the order fall within the trading session, is the news calendar covered, is there room within the account rules and the drawdown, does it not overlap with an open position, and does the size fit within the risk budget. Only when everything is green is the order recorded with strategy version, source version and a unique order ID, and then executed.
The role of ChatGPT and Claude
The models help with setting up research, code and rule review, analysing deviations and reporting. They only get task-specific, cleaned input and a fixed output format. A model response counts as unauthorised input until fixed validation accepts it. Models do not place orders, do not raise the risk budget and do not put their own code live. If a model drops out, the system keeps running without it.
Learning without the system giving itself more room
BH28 learns from its own results, within fixed limits. An improvement becomes a new version that is tested alongside the active version, against criteria set in advance. The active version stays frozen during that test. Failed variants are recorded too, so that afterwards it is not only the best-looking outcome that remains. And no single step can widen its own risk limits or permissions.
Where it stands
BH28 is complete and runs autonomously in simulation, without real money. A funded account only comes after a separate decision, once the simulation has shown for long enough that the chain holds: no duplicate orders, protection after every fill, and reconciliation with the broker without discrepancies. Software that works correctly does not yet prove a trading edge. That is why there are no returns here.
What this means for your business
With AI in your business, one question always comes up: what if it does something nobody approved? BH28 is my answer to that, in an environment where every mistake costs money straight away. AI supports, fixed rules decide, and every step can be traced back. I use the same building blocks for client processes with financial consequences, such as invoices, payments and stock: a control gate before every action, a log of every decision and a person at every exception.
Curious where that fits in your business? Book a free assessment.
Frequently asked questions about BH28
Does BH28 trade with real money?
No. BH28 runs autonomously on a simulation account. Real money or a funded account only comes after a separate decision, once the simulation shows that the chain is reliable. I do not claim returns, because working software does not yet prove a trading edge.
Does the AI in BH28 decide on orders?
No. ChatGPT and Claude help with review, analysis and reporting. An order only comes from the fixed chain of strategy and risk gate. A model response counts as unauthorised input until fixed validation accepts it, and without a model the system simply keeps running.
What technology does BH28 use?
A Python core for strategy and risk rules, 32 n8n workflows for scheduling, handover, monitoring and reporting, PostgreSQL for storage and audit, and ChatGPT and Claude as bounded helpers for review and analysis.
What does my business gain from this approach?
The same building blocks work for any process with financial consequences, such as invoices, payments and stock: a control gate before every action, a log of every decision and a person at every exception. That way AI can help without making the decisions itself.